Box’s Big AI Leap: Aaron Levie on Agents & the Future of Work

Hello.

Yes. We need, like, a newspaper hostage. Okay, this is when this happened.

Well, it depends on if I can decouple the stock price from a more technical, actual issue that we're dealing with. And so I would say I'm less worried about the stock price moving because of Wall Street trading up or down for any particular reason, but very worried about the embedded issue that might be related to that. So, let's say sometimes it's self-caused, so we have to execute better in a particular area, and the stock price in those situations is just a symptom of the actual thing that you're dealing with, and it's your real-time KPI of that problem.

And then in the case of tariffs, the thing that scares the living heck out of me is actually the economic impact that this would have. So I don't really care about the stock price per se. I care about the health of the country economically. And then obviously there's lots of follow-through that would happen and ripple effect that would happen if, as of Tuesday afternoon, the tariffs that have been proposed, if those roll through, this is gonna be a total disaster. So hopefully, whenever this airs, we'll be able to look back on this and laugh and say, ah, it was just an incredible negotiation gambit and it worked out.

And we're all—

Yeah, yeah, yeah, of course. Yeah, Xi just gave up and it was just like, wow, America, you're so strong, and we're going to give in, and that'll be amazing. And I'll be the fastest person to turn and say, wow, good work. My fear is that doesn't happen. My fear is that there are certain countries and groups of countries that have more leverage than we realize, and that we've gotten ourselves into a negotiating position that is going to make it harder and harder to pull back from.

And Trump specifically is obviously in a situation where he doesn't want to be perceived as being weak. But you don't want that as your core lead negotiator in a situation when you might not have all the cards. And that would be the scary thing about this: do we run into a negotiation that we can't pull ourselves out of? And all of the fallout from that would be obviously very problematic.

Hmm.

Yeah, exactly. I mean, it's interesting that I have a couple of categories of issue, right? So, like, the cheap T-shirt and cheap shoes, we'll probably find a way to survive in terms of maybe we'll have less consumer goods on that front. That would be pretty bad for certain parts of the economy. That's not good. But maybe you just buy a little bit fewer things.

The part that I think is disastrous is, think about all of the input materials we need to be able to build our industrial sector, to build our companies that go in and then sell things all around the world. What you're basically doing is you're handing to our competitors an instant disadvantage that we now have, where we're basically saying, you know what, we want our input costs to be higher in the U.S. And now you have entrepreneurial people globally saying, well, this is great.

Like, the U.S. just erected its own barrier to being able to compete with us. What an incredible moment, right? If you're an international entrepreneur right now and you're watching this, you're like, I can now finally outcompete the U.S. because they're just going to create friction for their own businesses. So there's actually no good story here. The only good story would be some scenario where you'd done this, like, 100 years ago and we just miraculously had robotic factories at this point because we had made this decision.

But we're not on that timeline.

Yeah, exactly. Yeah, right, right.

And so we have to deal with the timeline that we're actually on. The timeline that we're actually on is one where we have an interconnected supply chain globally, and our companies need that to function and to have some degree of predictability to be able to operate in this economy. So that's why I think we're running off a cliff if we don't pull things back. The one saving grace that I'm optimistic about on this path is Elon caring about this issue. He's already been pretty outspoken about it in a couple of ways, some subtle, some much more direct.

Kind of against the head of trade. And by the time this airs, something will have happened between that dynamic. It's not possible for it to sustain in this way. So we'll see what happens. And I'm hoping that Elon bails us out, or maybe Scott Bessent wakes up to how big of an issue this is. Something will happen between now and then.

I would say it's probably, for us where we're at, neutral to positive. But that's idiosyncratic to us. We sell to enterprises. We're at a stable part of the growth curve where it's not like we're either going to grow 50% or 17% next year. We kind of have bands of predictability, which means that when you have predictability, you can kind of know what your cost structure should be. You sort of know how to guide to Wall Street.

So for those reasons, let's say those are neutral-to-positive dimensions. There's this one element, which is, okay, you're kind of pivoting the business in real time while being public. I think that it depends on how costly the pivot is and how much upside versus downside is there in the near term. Those are the times where pivots become very challenging as a public company. I think in the case of AI, it's somewhat costly for us in the sense of we have to put a lot of engineers on this problem.

It's not so costly from a compute standpoint because we've sort of been able to price our software in a way that can support the actual compute of the AI workload. So it's costly from R&D. There's some heavy change management internally to execute on the opportunity. But that would happen whether you're public or private, so that's not really that different. And then I think for us, our business model is probably, if anything, enhanced because of AI, because now the use cases we can solve for customers are more significant.

We can actually increase the market opportunity that we're going after. We can get into the implications of that. But we think the TAM of our category goes up because of AI. And the business model doesn't get severely disrupted in the sense of we used to be selling seats and now all of a sudden something totally different. It's seats and then something in addition to that. So if you contrast that, let's say, to maybe Blockbuster, right, going through—if you were to go back 20 years and say, "Blockbuster, would you like to be a public company or a private company while competing with Netflix?"

You'd say private, because what you do, you have to take down the business model of the late fees. You have to guide Wall Street in a very different way. And then you have to be like, "We are going to have a different business model. It's going to be worse on some dimensions. It's going to be better on others. We're not going to have as much capital expenditure on real estate, and the model is different, but we're going to take some hits."

Those are the times where it's easier to be private during the pivot. But for a lot of enterprise software in particular, it's mostly upside. And I think most companies are navigating this well. Google is an interesting one as an example. Sundar is doing an incredible job, but you can kind of feel for the challenge that he has, because if all of a sudden search dramatically changes to AI chat and the ad model is different, at least in enough ways that change the monetization, he's got to navigate that with hundreds of billions in revenue.

So that's actually maybe a harder challenge than the average software company that just sees this as pure upside. At least in their case, they've got these really interesting new revenue streams, AI workloads with compute, the TPUs, and that gives them some upside in the process as well.

Doesn't keep me up at night. Okay.

I don't think it's a problem. I mean, probably if I'm really selfish, it just means more shareholders for us. Like, if you just on a continuum said, would you like there to be 10,000 public tech companies or 1,000? Where would you want to be on that continuum? You can just instantly imagine the liquidity dynamics that would emerge. So I'm kind of neutral. I don't really care that much. I think the interesting innovation that has emerged is—and this is funny because people kind of don't know exactly the root cause.

Like, maybe people think 20 years ago we made it so hard to go public. I'm actually not in that camp. I don't think we made it too hard to go public. I think actually it's good to have a heightened degree of scrutiny and regulatory pressure on being a public company. It is absolutely net positive for the average shareholder that we have all of these systems and governance controls in place. I think that's only a good thing.

But no matter how we got here, where we're at is we now have this new innovation, which is late-stage capital that can basically keep you private for, as far as we can tell, maybe forever, because you can kind of outrun the problem of converting these burn companies to cash-flow-positive companies. And usually you couldn't outrun that in the private market, so you ran out of funds and had to go public to get that capital. In a world where you can just keep raising privately, you can then twist the business model at some point to generating cash.

And you can kind of bend your mind a little bit and be like, well, at the end of the day, if you have a flow of cash coming into the company, you basically can make your own market for your shares, which means that any shareholder can basically decide at any point, do I like the price now or do I want to hold it? And you don't really have to—there's a very real scenario where a lot of these companies that are doing this model don't have to go public ever.

And there's not a premium that they give up in that process. In fact, there might be a premium to staying private because they don't have the volatility that we all deal with in the public market and so on.

So now you have this interesting issue where—okay, so if you play it out really far and let's say you had 30 really good companies doing this, it would be kind of funny because you might be like, well, why does that company have to go public? That I would actually say would be maybe a bad scenario because I do think just having choice in capital markets is probably good. But, for instance, the adverse selection that we saw, which was the micro version of that, is SPACs were basically adverse selection because you're like, okay, who needs to really go public right now?

Or who's just going public because it's convenient? Those are not reasons to go public. If you're going public because it's convenient and you're going public because you need to, those are instant filters for, like, don't be public. The kind of companies that should go public are like, you have predictable revenue, you're at a scale where that predictability will probably sustain. There's not a major disruption coming that was going to flip your model overnight. You have some degree of 1,000 employees, enough scale where things aren't going to blow up because one person leaves or this org change happens.

And so you don't really want companies that are rushing to go public before they have a lot of those conditions met. And so it would be interesting if there's a 10-years-from-now version of that, which is like all the numbers are five times bigger, but it's still problematic relative to the private companies. But I think if you were to design a system from scratch, we never even had as liquid markets as we have today, which is obviously a great thing and a great asset for the U.S.

I think you would just say, okay, I start a business, people can invest in that business, I can buy out their shares as the company is more successful. Warren Buffett really understands that model, and that would just be a model you could sustain forever if you wanted.

Hasn't changed.

So we started the company in college, and my co-founder and I, we were sophomores in college. And we had a variety of friends. And the reason I'm bringing them up is because we eventually had kind of four co-founders, but a variety of friends through high school. We'd all tried different projects and startup ideas together. So we actually all went to high school together. A few of us went to middle school together. And so we'd known each other already, by sophomore year of college, 10 years, which is incredible. Incredibly lucky to be able to have a friend group that you've known for 10 years that you're still in touch with and you can start companies with in college.

And so I was studying. This was a time back in 2004 where it was weirdly hard to move data around. You had to have USB thumb drives. You sent yourself emails of file attachments. You had to have external hard drives. And it was a mix of schoolwork, a mix of—I had an internship which was using some legacy enterprise software, and I was doing a class where I had to develop a business model—or not a business model, a sort of a SWOT analysis on a market and then identify the companies that were in that market that had pros and cons.

And for some random reason, you never kind of know, like, why did the very initial thing pop in your head? That's the magic of entrepreneurship. I chose this category of online storage.

As one does in college. And it was obviously the weirdest—like, everyone's doing, "I'm going to choose Nike," and, "I'm going to choose Walmart." And I chose online storage, which is obviously totally the weirdest thing to do. But what was really cool is it gave me this sort of excuse to think about a new product area. I started brainstorming the idea with some other people.

Some of the other—the two other co-founders kind of included in that. And instantly it kind of hit me and this group that, okay, there maybe was an opportunity, which was storage was getting cheap, compute was getting cheap, internet was getting faster. People were hopping around more devices and more computers. And so maybe there was this moment where you could access your data from anywhere and you just put it on the internet. And it wasn't called the cloud, but you put it on the internet and you access it from any device.

And that was the original idea. So we launched it, again, first the first co-founder and myself, and then it didn't—

That was Dylan. So we launched it, and it was funny. So Dylan's our CFO today, and he runs kind of a core chunk of the business. But just like back then, it was just so funny because he was, I think we called him, our head of marketing and PR. And the guy—I love the guy—he doesn't have a marketing bone in his body. So we just, everybody was like, we need different roles for everybody.

And so we launched this thing, and it didn't take off, but five people signed up the first week or something, and then, like, 20 people signed up the next week. And I started talking to some of the users, and it was the first time after, like, five or six years of doing web software, web applications, and web projects, it was the first time I would actually talk to the human on the other end. I'd launched shitty stuff before where you just see traffic and people are clicking things and clicking an ad, and you're like, cool.

This is the first time users, I would interact with them, and they would be like, "I like this. I don't like this." And I was like, wow, this is maybe what it feels like to launch a real company. People have feedback, they ran into a bug, they have new feature ideas, and for some reason they trust that this is a real thing. Like, I don't know how this is happening. I'm just here in class on my BlackBerry responding to their email.

They think we're a company. Like, this is shocking. And so that happened. We actually spent the whole summer working on the company full-time from Seattle. And it just grew a little bit more and more. And we raised a little bit of angel investment.

Yeah, Mark Cuban was sort of the second part of this, but we raised $80,000 for a post-money of—I believe this is the post-money—it was either $240K or $320K, but the post-money was either $240K or $320K. We raised $80,000.

That round is—we closed that round a little bit.

So we closed that round up, and we hadn't even heard of the word dilution. We were like, wait a second, you're telling me somebody is going to give us $80,000 and we're going to have a valuation? This is incredible. They could have probably bought half the company. This was before Shark Tank and nobody was doing blogs about, like, Paul Graham. Nobody was reading PG at the time. And so you didn't know about valuations and dilution and all this kind of stuff.

So people were going to hand us money for a part of our company. Like, wow, we made it.

No. So we had pitched everybody in Seattle. So we were at home in Seattle. We pitched literally everybody. So if you were a VC in the Seattle area in 2005, you got an email from me, or at least you passed it along to your associate and they didn't take a meeting with us. And we had like three or four meetings, rejected by everybody. Great process. I'm glad we got the rejections.

It built this kind of grit in us very early on. We got super lucky that there was this guy who basically, I think, saw like, okay, here's two kids, I'll help them out. It felt like a little bit more philanthropic, as opposed to like, this is going to be the real way that we make it big. And so he brought in three friends to invest. And so it was $20K each they invested.

That was great of them. Unfortunately, a slight asterisk on the story: they actually wanted to get bought back a few years later. So it doesn't have the happy ending on the multiple that could have happened. But, again, it was all fair. But they were fantastic on betting on the company early on. And we still just wanted to stay active talking to investors and people in the community. I had emailed Mark Cuban.

He got back to us in like 10 minutes and said, "Interesting company." We started working on a partnership, weirdly, in one of his companies, and then that sort of parlayed into an investment. So we ended up raising a few hundred thousand dollars in the early fall of 2005. And this is when we were back at college.

Yeah, for Mark.

Yeah. So Mark's prolific at this. There could be like a, I don't know why we don't have like a Discord at this point, but there's probably 2,000 of us or 5,000 or 10,000 people that have this exact story, and we're all the people that Mark saw in us and responded to some email. Funny enough, actually, then a part of the story—this, I don't think, has ever been in the version of it.

So maybe you're right that you have to kind of update the founding story every now and then. So as a part of the due diligence for the company, he had us meet Travis Kalanick. So he invested in Travis at this company called Red Swoosh. And so I had met Travis in L.A., and Travis was basically like, "Mark asked me to meet you, see if everything's above board." So, meet Travis. And I guess I got the thumbs-up.

And then actually, Travis ultimately was an inspiration for dropping out of college later. So later, four months later, as growth was kind of picking up, we started getting more and more users, and we decided that we had to kind of focus on the company full time. And so we decided to drop out of college, move to the Bay Area, and Travis was a major factor in that.

Well, so we were the—we started because of the cloud transformation, so it was a little bit more—

What's that?

We were the cloud. So we kind of rode that. No pivot was involved. We were not cloud from day one, running on the cloud, but we were cloud to the customer. And so probably actually the only sad thing was we actually had to build out infrastructure for the first decade and a half because literally we started the company six months too early. So we started the company, and our first launch was on these co-located servers.

And Amazon Web Services didn't launch until, like, maybe it was actually even a year later. But basically, we were off by, like, a year, year and a half to AWS kind of existing. So if we just started a little bit later, we would fully be able to ride the cloud trend on the infrastructure side. But we had to manage it all. So we were selling the idea of cloud to many companies. And I think that maybe the biggest difference today that I think about versus the AI wave is, in cloud, you had to spend—there was probably 10 years of convincing companies to trust the cloud and that they could relieve themselves of managing data centers and servers and the software themselves.

And they could trust moving that into a cloud environment. And you would go to banks, and it was just basically, like, a nonstarter. Like, the only meeting was to see if you would ever do an on-prem version. They weren't there to meet with you about your cloud version. It was like, oh, this is nifty. Like, when are you going to have the on-prem version? And maybe there were some scenarios where you could kind of slip in through a crack in the organization, some kind of test environment for pre-production use cases.

That was it. That was like 10 years of cloud, with one asterisk, because big credit goes to a lot of early CIOs that were very thoughtful about this. They saw the writing on the wall on the economies of scale you get from cloud, the better software you'd probably get. So there are plenty of CIOs out there that bet early, and we wouldn't exist without them. A lot of software companies wouldn't.

But by and large, cloud was sort of this thing you kind of resisted, slash reluctantly started adopting. AI, on the other hand, definitely has some similarities on the change management, the policies you have to set up, the compliance. Like, those are all very similar. But the difference is the energy is very different, where I think most companies I meet with—and I happen to be in New York and meeting lots of banks—the energy is absolutely in the category of, how many use cases can I apply this to?

How do I start to lean in more? We've got these 20 experiments running. We'd love to make these ones in production, those ones not. And there's no resistance. The only resistance is purely on a technical basis. Like, it won't work for that use case. It will work for that use case. It won't work for that thing. We can deploy it there. But there's no sort of philosophical resistance that you saw with cloud. So you're not convincing people about, hey, this AI thing is a very big deal.

Everybody is on board. They get it. They're seeing it in their personal lives. They understand why this is going to be impactful. So then it's actually—the change management is purely human-based, process change management, and then finding the right workflows and workloads where AI is actually going to deliver the ROI people want.

Yeah. Without ChatGPT, none of this would have happened. Now, maybe in a different timeline, somebody else would have done ChatGPT and Gemini would have launched first and whatnot. So we could still have been on this timeline. I think somebody would have figured out, like, hey, why don't we put GPT-3 into a chat wrapper? But it could not have happened in a better sequence, that it just exploded. Consumers got it. The next-generation workforce is fully addicted to this stuff.

So you actually have two really interesting pressures that are—I don't know if they're totally unusual, but this didn't happen in cloud. Maybe it happened in mobile, maybe it happened in PCs. Maybe it happened in the internet, actually. I just wasn't in a corporate environment to kind of feel it. But you have, like, the CEO is sort of like, hey, what are we doing about AI? And then you have the new workforce coming in, and they're like, I don't know how to work.

Like, if I don't have, like, a thing that can help me answer questions and autocomplete stuff and do this, like, I can't believe you guys type all these words out. Like, what are you doing with your time? Like, it will actually—I'm going through a journey mentally of this whole thing. Like, I have a couple areas where I'm just like, I have to type the whole thing for my brain just to make sure my brain is still active.

So, like, I write these emails to the whole company, and they're way too long. I don't even know if anybody reads them. They're, like, two and a half pages' worth of just, like, here's our strategy, here's what we're gonna do. And I basically say, I'm not gonna even look at AI because I just wanna make sure I'm still tapped in. But short of that, except for that email, I think we could easily look back 20 years from now, however people are working, and we're like, can you believe that you would sit at a computer and just think for, like, five days and then write out a report that said, like, we should launch a new device in Japan and that device will help us?

And it's like, come on. Like, we know why we're going to launch the device in Japan. It's just like, we just have to decide, like, what's the entry strategy? What are the numbers? And, like, why would we spend a week typing that out? Like, did we really improve our critical thinking the 500th time we did that? Or was that now just kind of baked into us? And so I have some nostalgia for probably what we're learning in ways that we can't predict, but then I don't have nostalgia for just the drain on human productivity of this drudgery.

So I think what's going to happen is you have a new workforce come in and look at us like we're crazy and like, I cannot believe you spend time doing that. Like, I just already prototyped the app. Like, in the meeting to brainstorm this thing, here's your app. We're just going to be like, holy crap. Now, lots of stuff will break, lots of things will have to change, but it will be a really interesting new way to work. So that's totally different from cloud.

Very few people had that kind of level of epiphany and sort of forceful change driven by the user population, as I think we'll see in AI.

So everybody has to get good at a few different things right now. One, I would say, is the kind of Crossing the Chasm framework. So Crossing the Chasm kind of breaks out different stages that technology kind of goes through. And the key is to think about technology. And in this case, the stages are like these really early adopters, then early—he calls them these early pragmatists. And there's this sort of chasm, which is there's a lot of things that we as early adopters do that never make it mainstream.

And so we get all hyped up and we're like, oh my gosh, everyone's going to wear contact lenses that have computers in them. And we're pumped. And then 12 of them sell or something. And I don't mean to—maybe somebody has it working somewhere—but it clearly is not yet mass market. So apologies to offending any startup. So you have these early adopters that kind of confuse us because we get these really interesting early data points.

And the funny thing is, the early adopters are exactly what you need in all forms of technology adoption. It's just, unfortunately, they're unreliable as evidence of any other category. And you can't a priori know if that audience is going to be the one that scales things. So then you've got these early pragmatists, which are—this is like the actual person in a company that's like, okay, maybe I can adopt this thing for—give me some acceleration in this area.

And that's like your first indication that maybe things are taking off. They want a little bit of an early advantage. But these are still the kind of crazy people in the company. They're not like the CIO of Goldman Sachs, who's like—in this case, actually, the CTO of Goldman Sachs is actually this person. So I chose exactly the wrong company.

Was that Marco? Marco is exactly this. And so probably Marco is the person you want to talk to to get this wave. But the CIO of some—I'll just pick on something—the CEO of a German bank, let's just say, is not there. Okay. Sorry, Germany.

Okay. Thank God for not saying French bank.

And so that person's going to only go when Marco goes. Okay, the CIO or the CTO of Goldman. And so you need to get these technology curves right. But the key is to not think about these on the broad basis of AI. That's too macro of a concept. You have to basically think about it per category. So what we probably need to do is kind of break out where this technology is in these curves per space.

So, okay, AI coding agents—well, sorry, AI coding, and then maybe we can add agents into that. AI coding is very clearly in the pragmatists. GitHub Copilot, now, I don't know, three years old. Cursor's flying off the shelves, Windsurf is flying off the shelf, Replit, all these guys are kind of being used everywhere. So you're in that kind of hypergrowth cycle of that early pragmatist to the pragmatist. And so that one's totally safe. Like, there's no—you can just extrapolate at this point.

Like, we don't have anything in our way. It's probably, if anything, just pure sales-rep-constrained on how this could grow faster. Conversely, let's say an AI coding agent that goes and, like, you send it off on a task: build the whole thing. I'll come back in two days. It's built. That's like the really early innovators. These guys are the crazy ones. They're just testing it in some use cases. Okay, so you kind of have that. So you have to go category by category to figure out where we're at.

A RAG, a document RAG system for chatting with documents, we're in the ascent on the early pragmatist to the pragmatist side. An AI outbound sales rep: early adopter. So I think everything needs to kind of be looked at through this curve, and then you'll kind of figure out where we're at in that acceleration.

Yeah, so the way to think about Box is we're a platform that helps you store, share, manage, and collaborate on data. We have a layer of security, we have a layer of file permissions, we have a layer of data governance. So that's what we've been working on for basically two decades. What we added was a layer we call the AI platform, and it has all the plumbing that you would expect you would need to be able to work on content in an AI context.

And so what do you need to be able to do? Some of it, by the way, benefits directly from other use cases that we've had. So, for instance, for 15 years, you click on a document, you see it in your browser. Well, why is that possible? It's because we have a conversion engine that takes the Word doc, makes it into a PDF.js, and that's how you get it. Well, guess what?

In that process, you've extracted—we can easily do text extraction. We can then do the embeddings on that text extraction. So that's a service that largely was built out for a different reason that now we have as a standing start. We then put that into a vector database. So now we have a layer that does the embeddings for a certain selection of content. We're not doing this for everything. So far, it's too expensive.

We have hundreds of billions of files.

So we're figuring out what we want to do for the broad corpus. But we have a feature that lets you target a certain set of data and then effectively be able to do RAG on that data. Then we have a layer that is the model abstraction layer. So it connects to right now four or five major models, but you can theoretically think about it as a complete bring-your-own-model architecture. Then we have an AI Studio that lets you go in and create custom configurations of a model, a set of instructions, and a set of tools.

That's how you create basically, today, primitive agents that will become much more comprehensive over time. So that's our AI platform layer. And then we have a set of APIs that will let you tap into any of those capabilities in any application that you're building or that you're using and working with as an end user. Or you can come to Box directly to do this. So the use cases are: I'm on Box, I'm on a Word document, summarize the document. Great.

I'm on Box. I'm looking at a contract. Please provide a legal analysis of this contract. Okay, great. Default didn't work well. Claude 3.7 Sonnet, provide the legal analysis of this contract. Great. Boom. You get the answer. So those are the basic use cases. Now imagine doing that to any amount of data. Okay, I want to take all my HR information and I want to enable anybody to ask questions of that.

I want to take all my sales materials and enable any sales rep to ask questions of that. So that's where we have this capability called Hubs, which is basically instant RAG on any collection of data within the Box environment.

Literally, yeah. So we were actually building this pre-gen AI, but we got really lucky that the ChatGPT moment happened before the final lines of code were getting written. And so that let us say, oh, this is the obvious interface for delivering this, because what you have is, imagine the scenario of you go to a completely general-purpose chatbot and you say, what's our HR policy?

Okay, that one might be a little bit easier. But let's say you say, what is the last earnings response that we had, or earnings message that we had? The challenge with that is it's going to pull from all your emails, all your documents, all your Slack channels, and there's probably lots of things that actually have your earnings messages. And so what we were trying to do is basically solve this problem of if you go into a complete general-purpose thing that's doing RAG or federated index on everything, you're going to get a lot of stuff that's coming back that's maybe not the best answer for that problem.

Or a lot of customers would go and build their own RAG stack. And we said, well, we could just give you a Hub, and then that Hub will be on earnings data, that Hub will be on HR data, that Hub will be on sales data. And it has basically the authoritative set of content that you can then talk to.

Yeah, exactly. So we're cheating in two ways. The user contributing the data to the Hub is only going to contribute the most relevant and authoritative content. The user coming into the Hub is psychologically now focused in on a topic. And so we've eliminated 95% of the two problems you get, which is data sprawl and unconstrained, unbounded queries that somebody could have. We think it's smart, and we sort of lucked into it a little bit, and now we're just running with it.

And then just to show how we think about this, we don't really care if the customer comes to the Hub itself or just uses an API and embeds it somewhere else. So we imagine a world where anybody else building software says, okay, I need to pull in knowledge into this workflow, and I have some agentic workflow that needs some kind of data about the HR documents of this company. The market and that customer has a choice.

They can either redundantly store their data in 20 different places, which is kind of illogical. Or we can have an approach where agents can run around and talk to each other. And so in our case, we'll have an agent. We already literally have this in our API. You can say, here's the Hub ID, here's the agent that I want to use, and I want to talk to that Hub. So now anybody else that's developing can just say, okay, I want to pull data from the HR Hub or from the sales Hub on behalf of that user.

And then surface up an answer. And so we want to make sure this is available to any developer. We'll plug in any system. We don't need to own the interface per se.

Yeah, so I think we want to be the best place in the world for companies to manage their content. The new definition of that, the jobs to be done of that, is certainly you've got to be the best place where they can automate the workflow around that content, use AI to ask questions on the data, and streamline those workflows. So we have to provide more and more of those experiences. So you can kind of wrap your head around the idea of eventually you come to Box, there's a marketplace of agents that you can pull into your data, and you have a legal contract agent, and you have a sales insight agent.

And we do want to enable those kinds of use cases. But I'm a big student of just nailing ecosystems properly. And there's a famous Bill Joy quote, who basically said there are always going to be more developers outside of Sun Microsystems than there are in Sun Microsystems. This was somewhat an allusion to Java and other technologies. And so the implication is: only do things that get the benefit of the broader ecosystem.

Never be only reliant on your own innovation and your own ecosystem. And so our version of that for this is basically twofold. One, we know that there's going to be an incredible player doing just a deep vertical workflow in a space. We still want to be a great data repository for that workflow. And sure, we're going to compete in some areas in that category because we might say, okay, there's some benefit to a consolidated user experience here for certain ones.

But we're also totally pragmatic and know that there's going to be scenarios where a customer just says, no, I've already invested in a vertical software product. I want that agent to talk to the agent in Box, where another set of data is. So we are going to be extremely open and interoperable, but we will also deliver great experiences for our customers and make it easy for them to deploy these use cases simply because we have about 120,000 customers, and there's going to be plenty of them that are not going to discover that amazing vertical app.

And we do have to offer these types of solutions to our customers as well. And so we're going to let the ecosystem kind of build on Box. We're going to enable this functionality for customers to build. We're going to partner along the way and hopefully have a few different sets of modalities in terms of how this comes together.

Do they choose?

Yeah. So, as of literally today's announcement—yeah, exactly. I don't know if I can break any news, but I'll allude to certain things. So we officially support the Gemini family. We support a variety of Anthropic Claude models, and then support OpenAI and the GPTs. And then I would just point to the idea that there's other breakthrough models emerging from Meta, from xAI. And our intent is to, again, consistent with us being as open and interoperable as possible, support models from across the ecosystem.

And then a very obvious end state is: just give us the API key and we'll plug into wherever your model is and whatever it is, assuming we can understand the APIs in a very clear way. So that's basically how it works. We don't yet do any kind of predetermination of which model is better. We sort of have a default for our basic functions, and then we just let you swap out to create your own agent when you have any other kind of use case that you want.

Yes, for a couple of particular customers that we just haven't announced yet, but coming soon.

We considered it for 10 minutes, and we said, no effing way are we going to get in this war. So it's back to the Bill Joy thing. It's like, why would you even want the brain aneurysm of thinking, like, oh, do we have all the latest researchers in our company, and we're competing for that, and we need to buy $2 billion of compute for the next training run?

That is just a war that you want to ride the tailwinds of, not play in. Now, different question if we were Meta and Amazon and those guys. But the moment you're not those, you have to understand where you are in the ecosystem and then figure out how to ride that. So we even debated for more than 10 minutes fine-tuning stuff. And even then we were like, no, because there's just model breakthroughs every day.

Why even have anything internally that could cause us to confuse ourselves that our fine-tuned thing is really important? There's a lot of things that you can do where—and actually, this started to become an interview question that I had for people coming into the AI team. I was like, do you want to fine-tune a model? And if they said yes, I had to beat that out of them and make sure that they weren't deeply committed to it, because AI is happening so fast.

I can't have somebody who's so committed to one particular architecture right now. We need range of motion. We need to be able to flip out something instantaneously. Internally, we kind of have a mental model of: if anybody has a breakthrough in the model space that we have built scaffolding or plumbing around to mitigate the lack of that breakthrough previously, we should just kill our stuff and just adopt whatever that breakthrough is. And maybe not within 10 minutes, but certainly the next couple of sprints that we're planning for, we should be thinking, like, oh, we were doing some crazy optimization on images with some technology.

If the model can now do that, just get rid of your stuff and just keep moving up the stack. Because what you don't want to do is ever have this culture set in of, like, our thing is better because you just happen to have the people that built that thing be so committed to it that you're kind of missing the new breakthroughs that are actually happening in the space.

So there's lots of stuff you have to do. We have one of our top AI guys. We think we invented a thing we call ERAG that does this sort of RAG enhancement that is able to do a better job at entity extraction from a large set of chunks of data and embeddings. So we do a lot of stuff ourselves, but it's only in service of, again, kind of like optimizing around the models right now in areas that the models are not particularly good at, but nothing that we think is going to lock us in place, which is why I've just been against fine-tuning or training our own model.

Because, again, you can easily hypothesize what happens when you go and train a model as a smaller player. Okay, wow, announcing the Acme X model, and then you get, like, two days of press and excitement, and then literally a week later, obviously Google's going to have an open-source model that blows you out of the water. And so now you have to decide, okay, first of all, why would anybody then use your inferior model? So now you're done.

And then do you have to do another training run with some new clever enhancements? Do you just do that for the rest of your life? Doesn't make any sense.

But in our case, some people were saying, well, you have access to all of this amazing proprietary corporate data. And the thing I proposed in some of these questions was like, okay, so what is corporate data to you? Like, what is that? And everyone's like, you jump to business documents and, like, a financial plan. And it's like, well, guess what? We have customers in movie studios and talent agencies and publishing houses for books and PR agencies.

And so would we train a model on every script in the world? Would we train a model on every clinical research document in the world? At some point, actually, you start to add up and you're like, oh, that's actually what the large language models are doing. They're training on literally all of the tokens you need to be good in a horizontal enterprise context are the exact same tokens you need in any of the large parameter models that we now have.

I can see if there's a very particular thing of, like, okay, it's just this model for just banks that just deal with this equity bond, this bond thing, and there's only four sources of data on the planet, and it's proprietary and closed off to everybody, then yes, you'd have to do this. In our case, the data is just too vast that you'd have to cover. And so the core model providers are all very good at it.

Yes.

Yeah, so back to the whole kind of layer of how we think about an agent internally so far is, and these are the kind of core primitives: you've got access to data, you've got the user permissions, you have a set of tools in Box. The tools right now are things like crack open a document and feed the model chunks based on the user's query. Search is a tool we're working on.

So you have a set of tools, and then you have the model, and then we're adding kind of an agentic workflow layer as well. And in combination, you can just think about it as agents for content. And so whatever you think a human does with documents and content, we should have an agent layer that lets you build that exact thing. And so what's amazing for us, and this is where I get most excited about AI, is in our business, I'm going to make up all the stats.

None of the numbers are exactly real, but I would argue that probably for 95% of our data, people have things that they would love to be able to do with that data that they just don't ever do because it's too expensive, it's too time-consuming. They didn't think to do it. They don't have a person with that expertise to do it. And so basically it kind of means that, like, 95% of your data is underutilized in some way. Like, if you could have an unlimited army of lawyers review every contract and they could understand everything in all of those contracts, would you pull out insights that would let you have a better renewal cycle with your customers where you could upsell them in the right way, or you knew what kind of terms you had to optimize for?

If you could have an unlimited amount of compute applied to all of your sales materials, and right when a sales rep is going into a meeting, you get the best insights on what to go sell that customer, would you sell more, right? Would you be able to go and generate more revenue? And so think about all these things that we just don't do today because it's just simply too complicated, too expensive, that you never have happen.

So the thing that I get excited about with agents is, I'm certainly fine with the use cases which are, okay, we used to have a person doing X, now it's an agent. Okay, that was an interesting optimization. The really big impact is the expansion of what people can do with software and start to solve the use cases that we just never ended up prioritizing in most of these categories. So for us, again, that's read all my documents, pull out the data from the document, put it into a database, let me automate a workflow, pull out data from my contracts, give me insights into all of those contracts at once.

A future use case: do a deep research exercise on due diligence documents, find similarities, differences, risks across an entire set of due diligence for an M&A transaction. So all of these kind of things that you can just now drop agents in to go and do way more of your work, and then you wake up and your work is done for you and you're kind of on to the next thing. That's sort of how we think about it.

Yeah, we've seen every problem that anybody's ever heard about. So the more agents you have, obviously you're adding probabilistic upon probabilistic upon probabilistic. So good luck with what that could lead to. General search problems still exist. At the end of the day, a lot of AI will be dependent on search technology and the quality of the search index, the quality of the ranking. Most of our biggest ideas and biggest problems intersect with search. And so the moment that the AI finds the wrong thing, the entire path is dependent now on the wrong thing.

So it's going to just make a whole series of bad decisions because it found the wrong thing as the starting point. And we can just see this in our personal lives using AI. It was funny, we asked every AI search product for consumers—we just did this funny test. We said, if you just do a query as simple as, tell me the last five times that the Giants beat the Astros on a Tuesday.

Okay, so you've got a very complex problem in terms of the dataset that has to be combed through and then looked at from a logic standpoint and all that stuff. Every single AI provider gave a different answer. I don't even know which one was right because we didn't care. But every provider—ChatGPT, Grok, Gemini, and Claude—gave us different answers. So what does that tell you? We have a search problem, right? Like, one found an article that was the wrong article.

One found a database of all the games, but then did the wrong thing when querying that. So you can just imagine enterprise datasets have that exact same problem in spades. One thing that people have run into in early AI rollouts is, if you at a large company, 5,000-person company, just had the most amazing agent of all time deployed and it actually found everything it was looking for, and you gave that to everybody, you could within the first day have employees finding corporate secrets that they shouldn't have access to.

So things like data governance matter because we've shared a bunch of stuff to people internally that either we accidentally shared or we only shared because we realized that they'd never find it. And so it didn't matter that you shared that thing with X group, but it's like, no, no, actually, probably you only wanted half that group to see those things, but you just didn't think about it. And so every enterprise has all these types of problems. So you've got a search problem, you have a data governance problem, you have a sort of chaining-of-agents problem.

So this is the part that will take years to fully figure out and get right. And the more intelligent the model gets, the better. But the quality of the model is only as good as the data you're feeding it. So again, that's why it's back to the search problem. So we're seeing all of those things. So the things that I feel very comfortable telling customers to deploy are areas where the problem is still very task-centric.

The thing is not going off and doing 50 steps. It's doing a couple steps. Ideally, the more agentic you get, the more you want it to be a task where there's going to be a user able to review the thing at the end. And so there's almost like a functional relationship to the user's ability to review the thing and how much agenticness you can afford to have. And so, for instance, I can have deep research go and do a very complex thing if it's about a space I understand, because it has now saved me 10 hours of Googling, and I can just quickly be like, it's wrong there, it's wrong there, that's wrong.

But I get a lot of value out of the general trend of the data. If it's a space I don't know anything about, I'm kind of like, oh, I don't know if I'm all in yet on all the data. And so it's important if you can review the work. This is why you probably shouldn't vibe code into production yet. There's a really funny viral video of AI fixing code bugs.

And it's this guy. Did you see this one where he's trying to paint the little precise part of his car that's off-color a little bit? And he does that, but he has tape and he just has a hole carved out in the tape so he can kind of paint this thing. And he rips off the tape, and then the whole car part comes off.

And it's like, that's like fixing bugs with AI right now. Because it's like, I'm gonna fix that problem, and we're gonna destroy the whole code in the process. And I've done this. You'll be vibe coding with something, and you're like, can you just please fix that one button over there? And it's like, all of a sudden, the whole product gets warped and flipped upside down, and the whole sizing's off.

And so this is what we're living with. This is the state right now. So you have to be in a position where you can review the work, you can understand what it's doing. And so in our world, that's like, okay, if you ask a question around 100 documents, it's basically going to get that right at this point. If you ask it over 10,000 on 20 different topics, you should be in a position where you can review all that.

If you ask it to pull out data from a short document, it's going to basically be 100% accurate on that. If you ask it to pull out data from a 200-page document and you only give it one shot to do it, it's going to get 75% accuracy. So you kind of need to know where the technology is. And yes, we do help our customers with that journey. The only thing that we haven't done as well yet, frankly, and hopefully by the time this airs, we'll have done it better.

I do like the idea that what's interesting is a lot of this is really a function of compute, right? So if I can send a task enough times to a model, I should be in a position to review all the answers and then start to grade where there is correlation, do a voting mechanism. And so we're not doing that enough at Box, but I like that the customer can just say, okay, this task is so valuable to me that I'm willing to pay for 10 times the amount of usage of the AI to get back an answer that has three more nines of reliability.

I think it'll follow probably the same curve that REST APIs did 20, 25 years ago, 23 years ago or whatever. We all started building apps that, we were like, oh, we need to plug into the Flickr API. And then that started to grow, and they were like, well, then I should have an API. And then we had all these things. So I think it looks like it's following that kind of pattern.

God bless Anthropic for coming out with MCP early in this to kind of say, somebody needs to plant a flag in the ground to get everybody to focus their attention on something. There's emerging concepts of agent-to-agent frameworks of, well, how do we have handoffs where I understand what your agent is good at, you understand what my agent's good at, I know how to auth in and do all that.

I think MCP could probably consolidate a lot of that energy, potentially, but there are other competing, or at least complementary, frameworks emerging. But no matter what, because of how many smart people are attacking this problem, I think we could easily be two years from now where you just feel totally comfortable that all the software you use can talk to all the other software in an agentic way. And then I think the more gnarly part is when developers have to start to decide, okay, am I going to use the API directly for this, or am I going to hand it off to agents?

And that's the only thing I'm a little bit nervous about, because clarity of integration patterns can be so, so helpful for velocity. And you might have, what will you do when you have a new generation of vibe-coder developers that think the world is just MCPs all talking to each other? And then they're like, okay, what we want to do is we need to move data from the Segment API to the Salesforce thing to that thing.

And they're like, great, we'll just MCP all this stuff. And then it's just like, everything's just 2% lossy in that transmission, or 20% more expensive than it needs to be because you're like, why are we running this through GPU when the APIs would have just been fine? So I do think that we're in for a little bit of needing to be like, okay, MCP and AI is like when people are talking to each other, and when the computers just need to talk to each other, you can still do the API thing.

We'll have to figure out where that goes.

I don't think there's any pushback. I think there's always a little bit of, especially with founders, that I get very excited about stuff.

It was. It didn't—well, it was founder mode of incubating the thing. It didn't require full founder mode to drive the full change because people got on very quickly. But you do have this risk, which is every one to two years as a founder, there's something that excites you. And then you have to kind of do this filtering process of, is this the thing that you're going to pivot the company over? And then the people around you can kind of see that, like, oh, you seem to get excited about this new stuff.

And I have a pattern where you get excited about something, but that thing didn't transform the business in the way you said. And so, especially if you have people that you've worked with for five or 10 years, that sort of can get embedded in. It's like, okay, but is this just like the same time you said VR is going to change everything? And so you have a little bit of this dynamic.

And I've even internally tried to frame things, which is like, okay, this is sort of not going to change the company, but it's still an important thing we have to hop onto. Like, okay, this is code red. It's going to change the company. And my conviction on that for AI happened very quickly. Like, the moment you saw ChatGPT, the moment you said, okay, I could put stuff into the context window and it's going to work on the data and come back with something.

Like, my first two days of ChatGPT, I started copy-and-pasting things from documents in and like, oh, would it respond well? And it was like, holy fuck. Like, this thing, the way you work with information will change forever. Now, I'm an idiot because I could have just been doing that in the GPT Playground, like, a month prior. But for some reason, it never dawned on me that in that text box you just paste a document and ask it to do things, so shame on me.

Somebody at Box was doing that, but I clearly had not had that epiphany in the same way. And so the ChatGPT interface, the moment I saw that, it was like, oh, this is going to change everything. And so through that winter, we started building a team, putting people together. I'm very lucky. We have a CTO from a company that we acquired who's just kind of the founder mode of CTOs. And so he jumped right on this as well.

And we kind of had the same wavelength on the topic. I'm lucky to have a broadly great leadership team. Our chief product officer jumped on it, our head of engineering jumped on it. So we had enough early people that said, okay, we're going to do this. It's going to be painful, it's going to be thrashy, but we're going to go do it. And it was actually funny: the head of engineering, separate from the CTO, when I first came to him, I said, okay, we need to allocate a team on this thing, but I promise it's going to be pretty limited because they just have APIs that we can talk to, and it's just plugging APIs into our platform.

And now there's a little bit of a running joke, which is he'll just say, like, it was just a few APIs, because now our AI team is probably the biggest team—yeah, easily the biggest team in the whole company in terms of any one group. And it's just because obviously the compounding nature of the problem just keeps growing.

So the cool thing is that AI has become a dominant enough thing that it's no longer, okay, this is that team working on the super experimental stuff.

Yeah, everybody now has a role to play. Every sales rep needs to become an expert on AI. Every support rep needs to become an expert on AI. Every engineer is in some way aiding our AI problem. It's taking over the end-user interface. The entire stack is doing something to help either the permissions model that serves AI or the embeddings infrastructure. So I don't think there's that many people at Box that don't think that they're doing something AI-oriented at this point.

But that took at least a year to a year and a half for that to fully become clear.

Yeah. So we have told everybody in the company that we want to use AI to be as productive as possible and aggressively use it across the business, with an asterisk on certain production use cases and whatnot. And we do internal things where we have members of the team show up at our internal all-hands event, which is a weekly video call, show what they're using AI for within Box, how it's helping them.

So we're trying to get everybody's creative juices flowing. I thought Tobi's memo was fantastic. It hit on basically the core topics you need to start to think about as a company. Here's how we should start to experiment. Here's how we should do product development. Here's how you should be thinking about when this can augment how we work and move faster.

So I thought it was a great conversation starter for a lot of folks.

So our stack is pretty flexible. By virtue of us doing so much in the data space, I would say probably by volume, Box AI is probably the number one thing, maybe in competition with GitHub Copilot. But Box AI, because we have a product called Box Notes where it's an open canvas to create any type of text-based content, we have all the models in there. So if you're generating anything, you can just choose your model and then spit it out in Notes.

Hubs for any kind of employee communication conversations. That's a lot of the AI usage. GitHub Copilot for coding—we finally got Claude in there, so that's got a lot of momentum. We're enabling Cursor for the VS Code crowd to be able to use. We've played with some of the AI agents, and I'm super excited about that space. Still pretty early for us, but we'll absolutely be using it. I like the idea of being able to deploy an agent to say, just make this SDK work in a new language, or there's a new Python update, please apply all the fixes.

That's such a great use case. I mean, just again, back to this category of the knowledge work that we waste in totally uninteresting, unstrategic areas. Nobody wants—it's like, I do. I'm very sensitive to the optics of the job issue. I'm actually not pessimistic at all. I'm extremely optimistic about the job issue. But as an example, there's just nobody, dude. If you looked at the kind of time we spend on—there's a new version of Python to fix a patch, and you have a team go dark for two months or something.

I'm making up all the numbers. So in case somebody says, well, you're an idiot for taking that long, it's a made-up scenario. But the team goes dark, and you're working on the most uninteresting stuff that the customer will literally never see. They will never know that you upgraded some library and all the dependencies you had to then fix and improve and everything. That is absolutely freaking insane that we spend so much time on that. Like, just have the AI do that, and then people should go think about the next cool breakthrough, get that working.

That's where we should be spending our time. Nobody's going to look back on that work and be like, man, it was so fun in my 20s and 30s being able to do bug patches for library upgrades.

Yeah, so we have pretty aggressive targets and goals that we've put out, where especially AI and AI agents provide us a number of new use cases that we can monetize. But I also think there's a counterpressure, which is that we should start to expect all of our software is intelligent and that you probably, like—I think you will grow faster because you're entering new markets, but you won't grow faster because there's some kind of magical AI premium on your software.

And those are two things to kind of decouple, which is, like, the price that you can charge because you have AI, that's not going up. And so, I think Wall Street probably got ahead of themselves for, like, three months two years ago, where they're like, "Oh, well, now you can just charge more because it seems cool." It's like, that was ill-fated as a concept. If you can go after new budget, then all of a sudden you can actually grow more and enter bigger markets.

That's going to be our case. But there's not some kind of magical, now you can double your price for your software type thing. It's just purely the use cases expand. So the next couple of years are basically getting into those use cases, having a TAM expansion. And I think there's going to be a lot of software that allows for that, whether it's brand-new startups going after spend that, again, was never kind of digital software spend before, or big companies that can go after more markets and opportunities as well.

And that's—I mean, I'm unbelievably excited for this moment in tech.

Hey, thanks for having me on. Appreciate it.

Bye.